FailPatch: Failure Residual Patching for Vision-Language-Action Models

2026-09-28

AI Summary

This study addresses the deployment vulnerabilities of Vision-Language-Action (VLA) models arising from insufficient supervision over failure states. To this end, we propose a failure-driven residual patching framework that decouples action generation from reliability supervision. By introducing a null-gated residual expert pool and a "retain-redirect-trust" mechanism, the method efficiently rectifies policies using failure trajectories. Built upon a frozen pretrained VLA, our approach achieves low-overhead adaptation with only 0.52% trainable parameters. Experimental results demonstrate that the proposed framework improves success rates by 11% on long-horizon tasks and 16.7% in real-world scenarios, offering an efficient and lightweight solution for the safe and reliable deployment of VLA models.

Abstract

Vision-Language-Action (VLA) policies are typically adapted using successful demonstrations, which provide direct action supervision but rarely cover failure-prone states. Deployment failures expose these states, yet lack the corrective actions needed for conventional supervised learning. We propose FailPatch, a failure-driven residual patching framework that decouples action supervision from execution-reliability supervision. Successful demonstrations ground how the policy should act, while deployment trajectories indicate when its behavior becomes unreliable. We further observe that action hidden representations exhibit clear linear separability between reliable and failure-associated states while directly conditioning action generation. Building on these insights, FailPatch introduces a Null-gated Residual Expert Bank into the action hidden space of a frozen VLA policy. A unified Preserve--Redirect--Trust objective retains the original policy in reliable states, selects residual experts in failure-associated states and redirects representations from failure regions toward success-associated regions under bounded intervention. With only 0.52% trainable parameters, FailPatch improves success rates by 11.0 percentage points on four long-horizon RoboTwin tasks under clean evaluation, 9.5 percentage points under clean-to-random generalization, and 16.7 percentage points over the baseline across three real-world tasks. Project and code: https://github.com/yupeng-2003/FailPatch.

Problem

• Vision-Language-Action Models
• Deployment Failures
• Failure-prone States
• Action Supervision

Innovation

• Failure-driven residual patching
• Vision-Language-Action models
• Null-gated Residual Expert Bank
• Hidden representation separability
• Parameter-efficient adaptation